💹 Forex Market Prediction Engine

A powerful, production-ready machine learning system for forecasting Forex prices using a stacked ensemble of Prophet, XGBoost, and LSTM models — with an auto-selected meta-model based on performance — deployed via a Flask web application that supports forecasting, plotting, trading (MT5), and trade log updates.

📌 Features

• Advanced Feature Engineering: Technical indicators, lag-based features, volatility, momentum, etc.
• Triple-Model Forecasting: Prophet (time-series), XGBoost (gradient boosting), LSTM (deep learning).
• Meta-Model Stacking: Best out of Ridge, Lasso, Random Forest, Gradient Boost selected via performance benchmarking.
• Quantile Prediction: Median, lower (10%), and upper (90%) quantile forecasts.
• Flask API: Routes for forecast, plotting, trading, and trade logging.
• Live Trading Enabled: Connected to MetaTrader5 for executing trades.
• OANDA Support: API integrated and secured via .env.

🧠 Model Pipeline Overview

Raw Data ──▶ FeatureEngineer ──▶ Train Prophet/XGBoost/LSTM ──▶ Stack with Meta-Model ──▶ Save Quantile Models ──▶ Forecast + Trade

🧪 ML Stack

Model	Purpose
Prophet	Trend + Seasonality
XGBoost	Non-linear patterns
LSTM	Temporal dependencies

🧰 Tech Stack

• Python, Pandas, NumPy, Scikit-learn, Matplotlib
• Prophet, XGBoost, TensorFlow / Keras
• Flask, MetaTrader5, oandapyV20
• .env for environment secrets
• joblib for saving models

📂 Project Structure

├── run.py
├── models/
│   ├── prophet.pkl
│   ├── xgb.pkl
│   ├── lstm.h5
│   └── meta_model_rf.pkl
├── src/
│   ├── forecast.py
│   ├── features.py
│   ├── evaluate.py
│   ├── mt5.py
│   ├── utils.py
│   └── logger.py
├── templates/
├── .env
└── README.md

🌐 Flask API Routes

Route	Method	Description
/	GET	Health check.
/forecast	GET	Returns next 7-day forecast as JSON.
/forecast/plot	GET	Renders HTML plot with prediction & confidence.
/trade	GET/POST	Auto-trades based on model signals.
/update-trades	GET	Updates log of closed trades.

🔐 Environment Variables

Create a .env file with:

OANDA_API_KEY=your_oanda_api_key
MT5_LOGIN=your_mt5_login
MT5_PASSWORD=your_mt5_password
MT5_SERVER=your_mt5_server_name

🚀 How to Run

1. Clone the repo

   git clone https://github.com/SteveParadox/TF_Analysis.git
   

2. Install dependencies

   pip install -r requirements.txt

3. Create .env file (see above)

4. Run Flask app

   python run.py

📈 Forecast Sample Output

{
  "forecast": [
    {
      "Date": "2025-07-23",
      "Predicted_Close": 1.0965,
      "Lower_Band": 1.0930,
      "Upper_Band": 1.0999
    },
    ...
  ]
}

🧠 Model Training Summary

1. Feature Engineering

   • Technical indicators (RSI, MACD, Bollinger Bands)
   • Lag features, rolling statistics
   • Candlestick patterns

2. Base Models

   • Trained independently on engineered features.

3. Meta-Model Selection

   • Models: Ridge, Lasso, Random Forest, Gradient Boost
   • Best model selected based on lowest validation MAE.

4. Quantile Stacking

   • Three meta-models trained for 10%, 50%, 90% quantiles for confidence bands.

⚙️ Live Trading

• Uses MetaTrader5 API.
• Trades only on EURUSD symbol.
• Uses forecast quantiles to place cautious limit orders.
• Can be extended for more strategies & risk management.

🛠 Extending

You can easily extend the system to:

• Add more base models (LightGBM, CatBoost)
• Use alternative data (news, sentiment)
• Deploy via Docker / AWS / Streamlit
• Enable backtesting or Telegram alerts

🧾 License

MIT License © SteveParadox